Alias Analysis

1
Alias Analysis

• Team 6
• Rakhi Anand
• Phanindra Kanakamedala
• Raj Kumar Jagiripu
• Quinn Lewis
• Chang Yun

2
Alias Analysis Basics

• What?
• Determination of storage locations that may be
  accessed in two or more ways
• Why?
• Essential to performing most optimizations
  correctly
• When?
• Applied before other analyses
• Where?
• Generated in WOPT/IPA and used in IPA, LNO, CG

3
Project Objectives

• Understand alias analysis in the compiler
• Export alias analysis information from Open64
• Alias classification, flow-free and
  flow-sensitive pointer analysis
• WOPT and IPA
• Create tool that displays the alias analysis
  results using the compiler on a given program
• Test tool with different applications

4
Static Single Assignment (SSA) form

• SSA form - Only one definition allowed per
  variable over entire program
• Each definition of a variable is given a unique
  version
• Different versions of the same variable can be
  regarded as different program variables
• Original Representation
  SSA Representation

a0 3 a1 a0 2 b0 a1 3
a 3 a a 2 b a 3
5
Aliasing Conditions

• Aliasing of a scalar variable occurs in one of
  four conditions
• When its storage location partially overlaps
  another variable
• When it is pointed to by a pointer used in
  indirect memory operations
• When its address is passed in a procedure call
• When its a non-local variable that can be
  accessed from another procedure in a call or
  return

6
SSA with Aliasing

• Factoring Operator ( ) - All edges pass through
  it, when multiple edges cross a join point
• Two types of definitions of a variable that
  characterize the effects of aliasing
• MustDef - operator
• MustDef must redefine the variable
• MayUse - µ operator
• MayUse only potentially redefines the variable

7
SSA and Aliasing

• Aliases in real programs can be modeled
• completely and concisely in SSA form
• Both direct and indirect memory accesses can be
  represented uniformly in SSA form using global
  value numbering
• SSA-based optimizations on scalar variables can
  be extended to indirect variables
• Any construct representing data storage can be
  represented in SSA form and benefits from SSA
  based optimizations

8
Alias Analysis in IPA

• One of Interprocedural Analysis phases
• Performs on the High WHIRL
• Computes all possible aliases in a program using
  information passed across procedure boundaries
• Objective
• Propagate alias information into procedures
• Propagate aliasing effects of the procedure out
  to the caller

9
Dataflow Relationship Between Modules
10
Five Phases of IPA
Phase 1 Inliner Phase (INL)
Phase 2 IPA Local Summary Phase (IPL)
Phase 3 IPA Analysis Phase (IPA)
Phase 4 IPA Optimization Phase (IPO)
Phase 5 IPA Miscellaneous Phase (IPM)
11
IPA Modular Hierarchy
ipa_driver (ipc_main.cxx)
Perform_Interprocedural_Optimization
(ipo_main.cxx)
Perform_Interprocedural_Analysis (ipa_main.cxx)
Padding Analysis (ipa_pda.cxx)
Optimize_Global_Variables (ipc_pic.cxx)
IPAA Simple IPAA
Eliminate_Dead_Function (ipc_cg.cxx)
IPAA Direct References Analysis
Do_Simple_IPAA (ipaa.cxx)
IPAA Indirect References Analysis
Clone_df (ipa_main.cxx)
IPAA Global Mod/Ref Listing
Ipa_Cprop (ipa_cprop.cxx)
IPAA Formal Points-To Mapping
Array_df (ipa_main.cxx)
Perform_Inline_analysis (ipa_inline.cxx)
12
Flow insensitive and sensitive

• Flow-free analysis (Compute_FFA)
• Creates the initial MU and CHI lists
• Initializes the POINTS_TO data
• Generates the occurrence table and virtual
  symbols
• Flow-sensitive analysis (Compute_FSA)
• Updates the POINTS_TO values already determined
  during flow free analysis
• Requires SSA representation
• Uses def-use (DU) chains to improve the
  information obtained during FFA

13
ALIAS_MANAGER Data Structure

• ALIAS_CONTEXT
• Determines rules from compiler options
• And dynamically from context (e.g. multi-language
  inlining)
• ALIAS_RULE
• 6 main language independent rules
• A few special rules for C and FORTRAN (no C)
• Exposed via Aliased() function in
  ALIAS_MANAGER
• ALIAS_RESULT
• NOT_ALIASED two memops are not aliased
• POSSIBLY_ALIASED cant prove memops are not
  aliased
• SAME_LOCATION memops are aliased and have the
  same memory locations
• POINTS_TO
• Summarizes the alias information about memory
  operations
• Each memory operation has its own POINTS_TO
  information

14
ALIAS_CLASSIFICATION

• Actually two C classes
• ALIAS_CLASSIFICATION
• IP_ALIAS_CLASSIFICATION
• Same alias class if
• Two expressions or variables may be aliased
• Two variables or expressions may be pointed to by
  the same pointer
• This is the biggest conservatism inherent in the
  algorithm

15
Points-to versus Alias Analysis

• Points-to analysis gathers the information
• Flow free
• Flow sensitive
• Alias analysis uses the information
• Symbol to symbol
• Memop to symbol
• Memop to memop
• Call to symbol
• Call to memop

16
Test Cases

• Purpose
• Generate data for the alias analysis tool
• Study the effects of the different compiler
  options
• Problems
• Very little change in the generated WHIRL code
  with small test cases
• Optimizer too good making it difficult to
  identify aliasing
• Dead code elimination
• Subexpression elimination

17
Compiler Options

• uhcc -O2 OPTaliasany,typed,unnamed,restrict,di
  sjoint
• any any pair of memory references may be
  aliased (the default)
• typed pointers to distinct data types are not
  aliased
• unnamed assume pointers never point to global
  variables
• restrict different pointers never point to same
  memory location
• disjoint assume multiple pointer indirection
  never overlap

18
Example of aliasing (C code)

• test.c
• include ltstdio.hgt
• int main()
• float arand()
• float b, c
• if (a1)
• b a
• else
• c b
• printf("\n u u u", a, b, c)
• return 0

19
Example of aliasing (WHIRL code)

• LOC 1 6 float b, c
• LOC 1 7 if (a1)
• LOC 1 8 b a
• U8LDA 0 lt2,1,agt Tlt39,anon_ptr.,8gt
• LOC 1 9 else
• LOC 1 10 c b
• U8U8LDID 0 lt2,2,bgt Tlt39,anon_ptr.,8gt alias_id
  12,fixed
• U8STID 267 lt1,5,.preg_U8gt Tlt9,.predef_U8,8gt b
  alias_id 1,fixed freq 0, ln 10, col 0
• U8U8LDID 267 lt1,5,.preg_U8gt Tlt9,.predef_U8,8gt
  b alias_id 1,fixed
• U8STID 266 lt1,5,.preg_U8gt Tlt9,.predef_U8,8gt c
  alias_id 1,fixed freq 0, ln 10, col 0
• LABEL L5 0 freq 0, ln 0, col 0
• LOC 1 11 printf("\n u u u", a, b, c)
• U8LDA 0 lt1,32,(11_bytes)_"\n_u_u_u\000"gt
  Tlt43,anon_ptr.,8gt
• U8PARM 2 Tlt37,anon_ptr.,8gt by_value
  alias_id 0
• U8LDA 0 lt2,1,agt Tlt39,anon_ptr.,8gt
• U8PARM 2 Tlt39,anon_ptr.,8gt by_value
  alias_id 0
• U8U8LDID 267 lt1,5,.preg_U8gt Tlt9,.predef_U8,8gt
  b alias_id 1,fixed
• U8PARM 2 Tlt39,anon_ptr.,8gt by_value
  alias_id 0
• U8U8LDID 266 lt1,5,.preg_U8gt Tlt9,.predef_U8,8gt
  c alias_id 1,fixed

Aliasing
20
Alias Analysis in Dragon tool

• Helps to understand and improve complier codes
• Prints essential program structures and obtain
  runtime information on their applications
• Consistent with other analysis options in Dragon,
  including Callgraph, Flowgraph, Array Data
  Dependence, Array Regions

21
Dragon File Import Internals

• All files are binary and generated with dragon
  compiler option

22
Aliasing